05 / Products
Solve/Agentic Hard-Problem Research
Open a hard problem. Let node agents search.
Point Solve at a Millennium statement, a Kaggle-style modelling question, or a precise conjecture. Each node owns an axiom session and a hidden agent. Start / Pause / Stop a run; the service ranks the frontier and wakes one agent per step. The kernel still will not QED an open Millennium conjecture.
Search
A run wakes one node agent; it expands, evaluates, and the next pick follows the score.
Start a run. The tree is the ledger.
Solve is an unattended research product for hard logical problems. A node is axioms, free-text reasoning, and an evaluating agent. Children inherit a forked copy of the axioms. Chat may narrate. Nothing stands unless it is on the tree.
It is not Stack. Stack investigates open-ended questions against organisational data. Solve searches programmes of attack: it gathers scholarly and web evidence, keeps mathematics as first-class markdown and MathML, and uses a logic kernel that will refuse an unearned QED.
A solver run is not a cron. Start, pause or stop it. Each step the service ranks open nodes by score and interest, then wakes that node’s agent to ideate, generate Horn consequences, or mix idea variants with a genetic algorithm the model grades. The next step chooses again.
A research chat that only talks will sound finished while the argument is still a transcript. Solve puts an agent on every node, scores the frontier, and lets a run expand what looks promising — so a person can pause, inspect, and see what was earned.
How a problem is worked
Open a claim. Wake node agents. Leave open what is not earned.
A problem is opened as a tree. Each node inherits a forked axiom session and its own agent. Start a run: the service ranks open nodes by score and interest, wakes one agent, and that agent ideates, Horn-generates, or GA-mixes a child — then evaluates it. The kernel will not stamp a QED the proof-state has not earned. Unattended is search. A human can pause.
- 01
Open
- 02
Node agents
- 03
Start run
- 04
Expand
- 05
Evaluate
- 06
Choose next
What a node agent can do
- 01
Per-node agents
Each tree node owns an axiom session, free-text reasoning, a score, and a hidden agent that may only expand that node.
- 02
Start, pause, stop
A solver run is a step budget, not a cron. The service ranks the frontier and wakes one node agent per step.
- 03
Choose how to expand
The woken agent ideates a child, Horn-generates consequences from this node’s axioms, or mixes ideas with a genetic algorithm it grades.
- 04
Evaluate, then choose again
Every new child is scored for promise and interest. High-scoring or interesting nodes are what the next step wakes.
- 05
Logic kernel
Classify claims, fork axiom sessions, generate consequences, and refuse a QED the proof-state has not earned.
- 06
Scholarly evidence
Search the literature and the web — Wikipedia, PubMed, OpenAlex, Crossref, Semantic Scholar and related sources — as evidence, not as theorems.
Use cases
Where Solve is used.
01
Open conjectures
The problem
A Millennium-class statement or an open conjecture needs a reviewable argument, not a confident paragraph in a chat.
How it is addressed
Solve keeps the claim on a tree of node agents: axioms, reasoning and a score. A run expands high-scoring programmes of attack and leaves the prize conjecture open.
02
Modelling claims
The problem
A Kaggle-style question is treated as a vibe — a feature seems important — rather than as a falsifiable claim.
How it is addressed
Record the modelling claim, gather evidence, accept or reject it, and keep the surviving argument on the ledger.
03
Proof sketching
The problem
A sketch looks finished in prose while the checker would still refuse the last step.
How it is addressed
The logic kernel holds proof-state. Unearned certainty is refused; accepted nodes carry a rationale a person can review.
04
Evolving approaches
The problem
Candidate lemmas and methods multiply in a transcript, and it is unclear which survived a fair comparison.
How it is addressed
A genetic algorithm mutates and grades variants against an explicit rubric, then promotes champions onto the tree.
Limits worth being clear about
Unattended search, not a proof of an open conjecture.
- 01
Solve does not prove open Millennium conjectures. Unattended search is still search; an accepted node is a reasoned verdict, not a theorem.
- 02
Literature and web search are evidence, not proofs. A citation can support a claim; it cannot close one on its own.
- 03
Genetic-algorithm descendants come from the evolution service. Chat does not invent a population by narrating one.
- 04
A human can pause or stop a run. The tree is a ledger a person can challenge, including what the system has left open.
- 05
The default Solve agent has no warehouse credentials. Organisational data investigation is Stack.
Enterprise deployment
These products are delivered as engagements rather than self-serve licences. We start with your problem, your data, and the systems you already run.